What Advanced DNA Testing Means, and What Distinguishes the Tiers

“Advanced DNA testing” is a marketing phrase, not a technical category, and in practice it covers three genuinely different measurements: genotyping arrays that read a fixed panel of roughly 600,000 to 1,000,000 pre-chosen sites, exome sequencing that reads the roughly 1 to 2 percent of the genome that codes for protein, and whole-genome sequencing that reads essentially everything. If you want the version that does not lock you into someone else’s site list, the answer is short-read whole-genome sequencing at 30x mean coverage or better. The data should be delivered to you as a CRAM or BAM alignment file plus a gVCF, with the reference build named explicitly. Everything else in this piece is about why that choice matters and how to check that what you received is real.

The three measurement types, and what each one can and cannot see

A genotyping array is a hybridization assay. Short oligonucleotide probes on a chip are designed for specific positions, your DNA binds, and a scanner reads fluorescence intensity that gets clustered into a genotype call at each designed position. This is the technology behind consumer ancestry kits, and it is cheap and accurate at what it does: common single-nucleotide variants, typed at a per-call error rate usually well under one percent. Its limitation is structural. An array cannot report a variant it was not designed to measure, so a rare or novel coding change in your genome simply does not exist in the output. Array files also include “no-call” positions where clustering failed, typically a few percent of the panel. Imputation, which fills gaps statistically from a reference haplotype panel, produces probabilistic guesses rather than measurements. Imputed genotypes at rare alleles are the single most common source of confident-sounding nonsense in consumer genetics.

Exome sequencing uses capture baits to pull down coding regions before sequencing, which concentrates reads where most known disease-associated coding variation sits. The tradeoff is uneven capture efficiency. GC-rich first exons, segmental duplications, and regions with high homology to pseudogenes are systematically under-covered, and the boundaries of the capture kit determine what you can analyze. Capture also introduces a distinctive false-positive profile: recurrent artifacts that appear across many unrelated exomes and can be filtered by comparing against an internal set of previously sequenced samples, which is far more effective than filtering on quality scores alone.¹

Whole-genome sequencing removes the target-selection step entirely. Coverage is more uniform than capture-based exomes, intronic and intergenic variation is visible, and structural variants and copy-number changes become tractable because you can look at read depth and discordant read pairs across the whole reference. It is the measurement that continues to be useful as interpretation improves, since reanalysis of your existing reads costs compute time rather than a new blood draw. The cost trajectory that made this practical came from sustained engineering on read length, parallelism, and reagent chemistry rather than any single breakthrough.²³

What to demand in the deliverable

The quality of a genome test is mostly the quality of what you can inspect afterward, so specify the files. At minimum you want the aligned reads, a gVCF rather than a bare VCF, and the pipeline provenance. CRAM is preferred over BAM for the reads, at roughly a third the size at equivalent information for lossless CRAM. Provenance means aligner and version, reference build, variant caller, and the exact command lines.

The gVCF distinction matters and is often glossed over. A VCF lists positions where you differ from the reference. A gVCF additionally records reference-confidence blocks, so when you look up a position and see no variant you can tell whether it was confidently reference or simply never covered. Without that, “no variant found” and “no data” are indistinguishable, which is the failure mode that produces false reassurance.

Reference build matters too. GRCh38 is the right default, and hg19/GRCh37 coordinates are still circulating in older annotation resources. A position reported without a build is unusable, because the same chromosome and offset means a different base in each. If your provider hands you GRCh37, expect to run a liftover and expect a small percentage of positions to fail or map ambiguously.

For quality, the numbers we check first start with mean coverage, where 30x is the working floor for confident diploid genotype calls from short reads. The others are the fraction of the callable genome at 10x or greater (above 95 percent for a good 30x run), the duplication rate from PCR amplification (under 10 percent), and the estimated contamination fraction (below 1 percent). Coverage is a Poisson-like process, so a 30x mean still leaves a meaningful tail of positions at 10x or below, which is why mean depth alone is a weak summary and the callable-fraction number is the one worth reading.⁴

The verification steps we would run on any genome we received

Before trusting any downstream interpretation, confirm the data is yours and internally consistent. These checks take under an hour of compute and catch the errors that matter most.

Start with samtools stats and mosdepth on the CRAM to reproduce coverage and duplication independently of the provider’s report. If the numbers disagree with the report, that is a conversation to have before anything else.

Then run peddy on the VCF. It reports genetic sex from X and Y coverage and X heterozygosity, checks ancestry against 1000 Genomes principal components, and flags relatedness and sample-swap signatures. Sample mix-ups and mislabeling are common enough in real sequencing operations that this is a standard step, and the method was designed precisely to catch swaps, contamination, and mismatched metadata quickly.⁵ A reported sex that disagrees with your own is the clearest possible sign of a handling error somewhere in the chain.

Next, sanity-check the variant counts. A single human genome at 30x on GRCh38 typically yields on the order of 4 to 5 million variants relative to the reference, with roughly 4 million SNVs and several hundred thousand short insertions and deletions, and the exact figure varies with ancestry because the reference is closer to some populations than others. A count far outside that range points at a pipeline problem rather than an unusual genome.⁶

Finally, treat any single interesting variant as provisional until you have looked at the reads. Load the region in IGV and check that the supporting reads are not all on one strand, not clustered at read ends, not sitting in a homopolymer run, and not carrying multiple other mismatches that suggest misalignment. Short-read callers are reliable on average and unreliable at specific hard loci, and interpretation of individual variants requires the kind of population-frequency and penetrance reasoning that has repeatedly surprised researchers who assumed a rare coding change was obviously causal.⁷ Anything you would want to act on medically belongs in front of a clinical geneticist or genetic counselor, ordered as a clinical-grade confirmatory test. It should not be settled from a research-grade file.

Where sequencing still cannot give you an answer

Short reads are about 150 bases, and that length sets hard limits. Repeat expansions, large structural rearrangements, and the highly homologous regions of the genome, including several pharmacogenes and the HLA locus, are poorly resolved by short-read data and often require long-read sequencing or targeted assays. Somatic variation, meaning mutations present in only a fraction of your cells, sits below the detection floor of standard 30x sequencing. The reason is that distinguishing a variant present in 1 percent of cells from a sequencing error requires either very deep coverage or molecular-barcoding methods that tag each original DNA fragment before amplification.⁸ Phasing, which tells you whether two variants sit on the same chromosome copy, is also incomplete without long reads or parental samples.

Sequencing tells you about the genome you were born with, which is a fixed input. It does not tell you what your cells are currently doing. The argument for pairing a genome with measurements that change over time rests on that gap. The genome is still the piece worth getting right once, because you only need to measure it once.

Questions people also ask

What are the three main types of DNA testing?

In the direct-to-consumer and research context, the three are genotyping arrays (fixed panels of common variants), exome sequencing (protein-coding regions via capture), and whole-genome sequencing. In forensics the categories are different again, built around short tandem repeat profiling, mitochondrial DNA, and Y-chromosome markers, though sequencing-based methods are steadily replacing capillary-electrophoresis STR typing there as well.⁹

Why do DNA tests say 99.9 and not 100?

Because these are statistical statements. They are not logical proofs. A paternity result expresses a probability of paternity given allele frequencies in a reference population. A genotype call from sequencing carries a phred-scaled quality that is a likelihood. That quality is not a certainty. A reported 99.9 percent means the alternative explanations have been driven to low probability, and no amount of additional data drives them to exactly zero.

Can a DNA test be 100 percent accurate?

No, and the interesting question is where the error sits. Per-base sequencing error, alignment error in repetitive regions, sample contamination, laboratory sample swaps, and mistakes in the annotation databases used for interpretation all contribute, and the last two are usually larger than the first.¹ Verification steps like sex checks and relatedness checks exist because handling errors are real and detectable.⁵

Which is better, 23andMe or AncestryDNA?

Both run genotyping arrays, so for ancestry the difference comes down to reference-panel composition and database size rather than measurement quality, and AncestryDNA’s larger match database favors genealogy while 23andMe historically reports more health-related interpretations. For anyone who wants to analyze their own data, neither is the right tool, because an array gives you a fixed site list and no reads to inspect.

How accurate is DNA testing for ancestry?

Continental-level assignment from arrays is highly reliable. Sub-continental and country-level percentages are estimates against a particular reference panel, they shift when the vendor updates that panel, and they should be read as ranges rather than fixed fractions.

Footnotes

1. Karin V. Fuentes Fajardo, David Adams, NISC Comparative Sequencing Program, et al. Detecting false-positive signals in exome sequencing. Human Mutation, 2012. https://doi.org/10.1002/humu.22033 ↩ ↩²

2. Jeffery A. Schloss, Richard A. Gibbs, Vinod B. Makhijani, et al. Cultivating DNA Sequencing Technology After the Human Genome Project. Annual Review of Genomics and Human Genetics, 2020. https://doi.org/10.1146/annurev-genom-111919-082433 ↩

3. Jeffery A Schloss. How to get genomes at one ten-thousandth the cost. Nature Biotechnology, 2008. https://doi.org/10.1038/nbt1008-1113 ↩

4. Rachel L Goldfeder, Dennis P Wall, Muin J Khoury, et al. Human Genome Sequencing at the Population Scale: A Primer on High-Throughput DNA Sequencing and Analysis. American Journal of Epidemiology, 2017. https://doi.org/10.1093/aje/kww224 ↩

5. Brent S. Pedersen, Aaron R. Quinlan. Who’s Who? Detecting and Resolving Sample Anomalies in Human DNA Sequencing Studies with Peddy. The American Journal of Human Genetics, 2017. https://doi.org/10.1016/j.ajhg.2017.01.017 ↩ ↩²

6. Tuuli Lappalainen, Alexandra J. Scott, Margot Brandt, et al. Genomic Analysis in the Age of Human Genome Sequencing. Cell, 2019. https://doi.org/10.1016/j.cell.2019.02.032 ↩

7. David B. Goldstein, Andrew Allen, Jonathan Keebler, et al. Sequencing studies in human genetics: design and interpretation. Nature Reviews Genetics, 2013. https://doi.org/10.1038/nrg3455 ↩

8. Michael W. Schmitt, Scott R. Kennedy, Jesse J. Salk, et al. Detection of ultra-rare mutations by next-generation sequencing. Proceedings of the National Academy of Sciences, 2012. https://doi.org/10.1073/pnas.1208715109 ↩

9. Peter de Knijff. From next generation sequencing to now generation sequencing in forensics. Forensic Science International: Genetics, 2019. https://doi.org/10.1016/j.fsigen.2018.10.017 ↩